PocketMatrix

On the desktop machine accessing graphic memory is much slower than accessing system memory therefore flip() will seem slower ( at least that used to be the case with the graphics memory)

On the other hand PocketPC doesn't have such concept meaning that framebuffer memory is basically the same as your main memory so flip() should run exactly as fast as blit from one surface to another.

Anyway, flip() on PPC runs at 5-6 ms which basically corresponds to memcpy() for 320x240x2 bytes.
GapiDraw uses prerotated images ( memory buffers) to match device screen layout.
This saves some time on the final blit but causes performance problems for people who want to access memory buffers directly ( either stuff has to be rotated (lock) or people have to know surface layout when accessing it directly.)

Another option would be to keep internal memory buffers ( surfaces) in the default unrorated state and just do rotated blit in flip() call.
It would not be much slower since rotated flip() takes around 7 ms ( as opposed to 5-6 ms for unrotated)- not much of a difference but saves a lot of trouble ( IMHO of course)

Indeed, it is not hard to come with a Rotated Blit that is as fast as a straight memcpy(). The process is limited by memory bandwidth, not CPU speed.

As Johan says, the reason to do so is to reduce tearing.

As fast as memcpy ?

Hmm, I would love to see that one ( and I am not being sarcastic here- I would really like to see implementation that is just as fast as straight memcpy.)

Anyway, when rotating you are limited by reading only 2 pixels at a time and have to perform bunch of shifts and logical operations ( admittedly , quite fast on ARM cpu) to be able to write to the destination using STMIA instructions.
In my experience, just the fact that you cannot use ldmia/stmia instructions when writing and reading will result in slowdown of about 30 %.
This all assuming 200 MHZ ARM 1100 CPU.